vec_math
Class LevenbergMarquardt

java.lang.Object
  vec_math.LevenbergMarquardt

public class LevenbergMarquardt
extends Object

Solves for a data model fit using Levenberg-Marquardt algorithm as in Num. Rec., 15.5, p683ff.

Nested Class Summary

(package private) static class	LevenbergMarquardt.LMSineModel solve for phase, amplitude and frequency of a sinusoid
static class	LevenbergMarquardt.LMSineTest
(package private) static class	LevenbergMarquardt.SinusModel
static class	LevenbergMarquardt.SinusTest We test with a simple sinus fit.

Field Summary

static double	DEFEPSILON Default improvement in chi-2 to induce terminating loop.
static int	DEFITER Default number of maximum iteration.
static double	DEFLAMBDA Default starting value for the 'fudge factor' boosting hessian.
static int	DEFTERMINATE Default number of maximum improvements, if epsilon is reached.
private double	epsilon Required improvement.
private GradientModel	fit The differentaible multidimensional data model fit, dimension as x.
private double	lambda0 Starting value for lambda.
private int	maxiter Maximum number of iterations.
private int	maxterm Maximum number of improvements.

Constructor Summary

LevenbergMarquardt()
Constructs a new LevenbergMarquardt with default iteration and improvement values.

LevenbergMarquardt(double lambda)
Constructs a new LevenbergMarquardt with default iteration and improvement values.

LevenbergMarquardt(double lambda, double epsilon)
Constructs a new LevenbergMarquardt with default iteration and improvement values.

LevenbergMarquardt(double lambda, double eps, int iter, int term)
Constructs a new LevenbergMarquardt with default iteration and improvement values.

LevenbergMarquardt(double lambda, int iter)
Constructs a new LevenbergMarquardt with default iteration and improvement values.

Method Summary

VectorG	improve(VectorG start) Iterates for a solution to the data model, starting from the given start vector.
void	setModel(GradientModel grad) Sets the model fit.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Field Detail

DEFITER

public static final int DEFITER

Default number of maximum iteration.

See Also:

Constant Field Values

DEFTERMINATE

public static final int DEFTERMINATE

Default number of maximum improvements, if epsilon is reached.

See Also:

Constant Field Values

DEFEPSILON

public static final double DEFEPSILON

Default improvement in chi-2 to induce terminating loop.

See Also:

Constant Field Values

DEFLAMBDA

public static final double DEFLAMBDA

Default starting value for the 'fudge factor' boosting hessian.

See Also:

Constant Field Values

maxiter

private int maxiter

Maximum number of iterations.

maxterm

private int maxterm

Maximum number of improvements.

epsilon

private double epsilon

Required improvement.

lambda0

private double lambda0

Starting value for lambda.

fit

private GradientModel fit

The differentaible multidimensional data model fit, dimension as x.

Constructor Detail

LevenbergMarquardt

public LevenbergMarquardt()

Constructs a new LevenbergMarquardt with default iteration and improvement values.

LevenbergMarquardt

public LevenbergMarquardt(double lambda)

Constructs a new LevenbergMarquardt with default iteration and improvement values.

LevenbergMarquardt

public LevenbergMarquardt(double lambda,
                          double epsilon)

Constructs a new LevenbergMarquardt with default iteration and improvement values.

LevenbergMarquardt

public LevenbergMarquardt(double lambda,
                          int iter)

Constructs a new LevenbergMarquardt with default iteration and improvement values.

LevenbergMarquardt

public LevenbergMarquardt(double lambda,
                          double eps,
                          int iter,
                          int term)

Constructs a new LevenbergMarquardt with default iteration and improvement values.

Method Detail

setModel

public void setModel(GradientModel grad)

Sets the model fit. The model must extend the abstract gradient model. If the gradient is not accessible, or a higher robustness is needed, try using a Simplex, like Amoeba.

improve

public VectorG improve(VectorG start)

Iterates for a solution to the data model, starting from the given start vector. The solution, if fed into the getPseudoHessian of the fit is the inverse of the covariance matrix.